API reference

App name recorded on the generation. A non-empty value takes precedence over X-Title.

App name when X-OpenRouter-Title is absent or empty.

An opaque conversation id (up to 256 characters) keying session affinity: requests carrying the same id prefer the endpoint that served the conversation, keeping its prompt cache warm. On /chat/completions the session_id body member outranks prompt_cache_key, which outranks this header. On /responses, prompt_cache_key outranks this header. The resolved id is forwarded to the serving endpoint as x-session-id.

Model id as author/slug, optionally suffixed :free, :nitro (fastest) or :floor (cheapest) or :exacto (measured tool-call reliability, requiring declared parameters). :online is refused with HTTP 400 because web search plugins are unsupported. provider/slug or provider/author/slug names the model that provider serves under that slug and routes to that provider alone; a catalogue id always takes precedence.

Fallback list, at most 20. The first id that can be routed serves the request.

The conversation, at least one message. A user message's content is a string or an array of text, image_url, video_url, audio_url, input_audio and file parts; a *_url part carries either a URL the endpoint fetches or a data: URL carrying the bytes. Each part is routed only to endpoints declaring an input for the modality it needs, in the source and format it carries, and an input_audio part reaches an endpoint that takes audio as audio_url rewritten as a data URL. A file part carries its document inline as file_data, or names one stored with /files as file_id, never both: the marketplace reads the stored bytes into the request before routing, so an endpoint is sent a data URL it declared it accepts and never an id, and a stored image reaches the model as an image_url part. A request may name at most 16 stored files, 32 MiB in all. An assistant message may carry a previous reply's reasoning_details, sent back exactly as received, so a provider that binds it to a signature or holds it encrypted can verify or replay the same sequence.

Answer with SSE chunks instead of one object.

false is a storage opt-out: vetted ZDR endpoints satisfy it without a provider parameter. true is unsupported because chat completions are not stored and is rejected with 400.

Enable or disable thinking where the endpoint declares this control. Dynamo endpoints receive it as chat_template_kwargs.enable_thinking only when that nested boolean is declared.

The final usage chunk does not depend on include_usage. Set include_tool_dispatch: true with stream: true and tools to receive named tool_call_dispatch events from qualified endpoints. This option is consumed by the marketplace and omitted upstream.

Sampling temperature.

Nucleus sampling mass in (0, 1]. Zero is rejected with 400, not replaced with a default.

Sample from the k likeliest tokens; -1 disables the filter where an endpoint declares it.

Floor on a token's probability relative to the likeliest one; forwarded when the endpoint declares min_p.

Accepted for compatibility; omitted from the upstream chat body.

Penalise repetition by count.

Penalise tokens already present.

Multiplicative repetition penalty in (0, 2]; forwarded when the endpoint declares repetition_penalty.

Ask for a reproducible sample where the engine offers one.

Up to eight sequences that end the generation.

Ceiling on generated tokens.

The OpenAI spelling of the same ceiling.

Per-token bias in [-100, 100], keyed by token id.

Return log probabilities.

How many alternatives per position.

{ "type": "text" }, { "type": "json_object" } or { "type": "json_schema", "json_schema": { … } }.

Routing marker: send it to be routed only to endpoints whose schema adherence has been verified.

Up to 1536 function tools the model may call. Names must match [a-zA-Z0-9_-]{1,128}.

none, auto, required, or one named function.

Allow more than one call per turn.

effort is translated to reasoning_effort when declared by an OpenAI-compatible endpoint; the shorthand takes precedence. max_tokens supplies the Anthropic thinking budget. The object itself is never forwarded.

The shorthand for reasoning.effort. A level an endpoint has been shown to refuse is filtered out of routing on its own, and the rest of the declaration stands.

Put reasoning on the reply where the model emits it.

How much answer to produce.

One choice per request.

An opaque end-user identifier, at most 256 characters.

The conversation this turn belongs to, keying session affinity. Outranks prompt_cache_key and the X-Session-Id header. The resolved id is forwarded to the serving endpoint as x-session-id.

A cache affinity hint used when session_id is absent. Outranks the X-Session-Id header and is carried to the serving endpoint as x-session-id, rather than as a body parameter.

Opaque members echoed on the generation record.

What the reply may carry. Naming image or audio asks for generated media in message.images or message.audio; absent takes the endpoint's own default, which for a text model is prose.

How a generated image is drawn: n, resolution, aspect_ratio, size, quality, output_format, background, output_compression and seed — the drawing members of the image contract. A chat turn resolves resolution and aspect_ratio against the endpoint that serves it, exactly as POST /images does.

The voice and container for generated speech, as { "voice": "alloy", "format": "wav" }. Both members are required: an engine asked for audio has to be told what to speak with and what to encode to, and the marketplace picks neither for you.

Provider eligibility, ordering and measured performance preferences.

Alias for sort.partition: fallback groups by model; sort orders across models. An explicit provider.sort.partition wins, followed by this alias, then the workspace partition. Other values receive HTTP 400 naming route.

Prompt transforms. ["middle-out"] asks for middle-out compression against the chosen endpoint's context window, applied after routing has picked that endpoint. The system message, the earliest user turn and the latest turns are kept, messages are dropped from the middle, and a tool call is never separated from its result. usage reports what was actually sent. [] declines compression, and so does an absent member. Any other element receives HTTP 400 naming the element and its value.

The same ask as transforms, written as a plugin. [{ "id": "context-compression" }] asks for middle-out compression, and an entry may carry configuration beside its id, which is accepted and not read. context-compression is the only served id, and any other, web included, receives HTTP 400 naming the id.

Usage accounting. { "include": true } puts usage on the reply, cost included.

An OpenAI chat.completion, plus provider and usage.cost in USD. With stream: true the body is text/event-stream instead.

The body did not parse, or a parameter is outside the range the API accepts. A file part that names both file_id and file_data, or a request that names more than 16 stored files, is refused too.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family. A file_id the account cannot read answers 404 as well, whether the file is another account's, was deleted or has passed its expires_at.

The upstream did not answer within the deadline.

The stored files the request names total more than 32 MiB.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503. A request that names a stored file is also refused 503, to be retried, while the key store that stored files are sealed under cannot be reached.

An opaque conversation id (up to 256 characters) keying session affinity: requests carrying the same id prefer the endpoint that served the conversation, keeping its prompt cache warm. On /chat/completions the session_id body member outranks prompt_cache_key, which outranks this header. On /responses, prompt_cache_key outranks this header. The resolved id is forwarded to the serving endpoint as x-session-id.

Model id as author/slug, optionally suffixed :free, :nitro (fastest) or :floor (cheapest) or :exacto (measured tool-call reliability, requiring declared parameters). :online is refused with HTTP 400 because web search plugins are unsupported. provider/slug or provider/author/slug names the model that provider serves under that slug and routes to that provider alone; a catalogue id always takes precedence.

Fallback list, at most 20. The first id that can be routed serves the request.

The text to continue.

Return text deltas as SSE, ending with usage and [DONE].

Sampling temperature.

Nucleus sampling mass in (0, 1]. Zero is rejected with 400, not replaced with a default.

Sample from the k likeliest tokens; -1 disables the filter where an endpoint declares it.

Penalise repetition by count.

Penalise tokens already present.

Multiplicative repetition penalty in (0, 2]; forwarded when the endpoint declares repetition_penalty.

Ask for a reproducible sample.

Sequences that end the generation.

Ceiling on generated tokens.

How many log probabilities to return per position.

Repeat the prompt before the completion.

Text to place after the completion.

An opaque end-user identifier.

Provider eligibility, ordering and measured performance preferences.

Alias for sort.partition: fallback groups by model; sort orders across models. An explicit provider.sort.partition wins, followed by this alias, then the workspace partition. Other values receive HTTP 400 naming route.

An OpenAI text_completion, plus provider and usage.cost.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

An opaque conversation id (up to 256 characters) keying session affinity: requests carrying the same id prefer the endpoint that served the conversation, keeping its prompt cache warm. On /chat/completions the session_id body member outranks prompt_cache_key, which outranks this header. On /responses, prompt_cache_key outranks this header. The resolved id is forwarded to the serving endpoint as x-session-id.

Fallback list, at most 20. The first id that can be routed serves the request.

Provider eligibility, ordering and measured performance preferences.

Alias for sort.partition: fallback groups by model; sort orders across models. An explicit provider.sort.partition wins, followed by this alias, then the workspace partition. Other values receive HTTP 400 naming route.

Prompt transforms. ["middle-out"] asks for middle-out compression against the chosen endpoint's context window, applied after routing has picked that endpoint. The system message, the earliest user turn and the latest turns are kept, messages are dropped from the middle, and a tool call is never separated from its result. usage reports what was actually sent. [] declines compression, and so does an absent member. Any other element receives HTTP 400 naming the element and its value.

The same ask as transforms, written as a plugin. [{ "id": "context-compression" }] asks for middle-out compression, and an entry may carry configuration beside its id, which is accepted and not read. context-compression is the only served id, and any other, web included, receives HTTP 400 naming the id.

A marketplace model id, the same as /chat/completions takes.

The conversation, at least one message. Each is a user or assistant turn whose content is a string or an array of text, image, tool_use and tool_result blocks. A system turn placed among them is accepted, and its text joins the leading system prompt rather than staying where it sits. An image or document block whose source is { "type": "file", "file_id": "…" } names a file stored with /files, which is read into the request before routing.

Ceiling on generated tokens (required).

A system prompt, as a string or an array of text blocks. The text of any system turn in messages follows it, a blank line apart, in one leading system message.

Opaque request metadata; accepted and ignored.

Up to eight sequences that end the generation.

Answer with an Anthropic event stream instead of one object.

Sampling temperature.

Nucleus sampling mass.

Sample from the k likeliest tokens.

Tools the model may call, each a name, optional description and input_schema. The client runs them: an entry whose type names one of Anthropic's hosted web search versions (web_search_20250305 and later) asks for a search the marketplace does not run and is refused with HTTP 400, while a function the client declares under the name web_search is its own.

How tools may be used: auto, any, none, or a named tool.

Extended-thinking config; accepted and ignored, since a routed model decides for itself.

An Anthropic message object, or a text/event-stream of Anthropic events when stream: true. A picture the model generated is an image content block beside the text, its source a base64 one carrying the bytes and their media_type — or a url one where the endpoint answered with a URL of its own. Streamed, the block arrives whole: one content_block_start carrying the picture, then its content_block_stop. The dialect's usage is a required member stating both counts and a client parsing it rejects an object with a member missing, so a count the endpoint never reported is sent as zero and named in usage.heimdall.unreported, which lists every count whose zero nothing measured; a count absent from that list is the endpoint's own figure, and the extension member is absent from a reply whose counts were both reported and names both where no usage document arrived. message_start names both of them, nothing being measured when a stream opens, and the final message_delta carries the verdict, with honoured beside its usage where the request carried members the endpoint's declarations judged. No block of a reply states a search: there is no server_tool_use or web_search_tool_result block, no citation on a text block and no usage.server_tool_use member, because the marketplace ran none, and a client that counts searches from those members counts none.

The body did not parse, or a parameter is outside the range the API accepts. A file part that names both file_id and file_data, or a request that names more than 16 stored files, is refused too. A tools entry whose type is a dated hosted web search such as web_search_20250305 is refused before inference, with the entry's position in the message. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

The key is missing, unknown, revoked, expired or past its spend limit. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

The account has no credit left to cover the request. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family. A file_id the account cannot read answers 404 as well, whether the file is another account's, was deleted or has passed its expires_at. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

The upstream did not answer within the deadline. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

The stored files the request names total more than 32 MiB. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

The reply carried generated audio, which this dialect has no content block for. Ask for the model on /chat/completions, whose message.audio carries it. After a stream has opened the same refusal arrives as an error event.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

The marketplace failed to serve the request. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

The endpoint that was routed to failed, and no other endpoint could serve the request. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503. A request that names a stored file is also refused 503, to be retried, while the key store that stored files are sealed under cannot be reached. A request with stream: true refused before its stream opened answers in AnthropicErrorEnvelope; every other refusal answers in ErrorEnvelope.

A marketplace model id.

The conversation to count, in the Messages shape.

A system prompt, counted alongside the messages.

Tool definitions, counted alongside the messages.

The estimated input token count.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

An opaque conversation id (up to 256 characters) keying session affinity: requests carrying the same id prefer the endpoint that served the conversation, keeping its prompt cache warm. On /chat/completions the session_id body member outranks prompt_cache_key, which outranks this header. On /responses, prompt_cache_key outranks this header. The resolved id is forwarded to the serving endpoint as x-session-id.

Fallback list, at most 20. The first id that can be routed serves the request.

Provider eligibility, ordering and measured performance preferences.

Alias for sort.partition: fallback groups by model; sort orders across models. An explicit provider.sort.partition wins, followed by this alias, then the workspace partition. Other values receive HTTP 400 naming route.

Prompt transforms. ["middle-out"] asks for middle-out compression against the chosen endpoint's context window, applied after routing has picked that endpoint. The system message, the earliest user turn and the latest turns are kept, messages are dropped from the middle, and a tool call is never separated from its result. usage reports what was actually sent. [] declines compression, and so does an absent member. Any other element receives HTTP 400 naming the element and its value.

The same ask as transforms, written as a plugin. [{ "id": "context-compression" }] asks for middle-out compression, and an entry may carry configuration beside its id, which is accepted and not read. context-compression is the only served id, and any other, web included, receives HTTP 400 naming the id.

A marketplace model id, the same as /chat/completions takes.

A cache affinity hint that outranks the X-Session-Id header. The resolved id reaches the serving endpoint as x-session-id.

The conversation: a string, or an array of message, function_call and function_call_output items. An input_file content block carries file_data or file_url, or names a file stored with /files as file_id, which is read into the request before routing.

A system prompt, prepended to the conversation.

Tools the model may call, each a flat { type: "function", name, description, parameters }.

How tools may be used: auto, none, required, or a named function.

Reasoning config; its effort maps to reasoning_effort.

Output shape. format is {"type":"text"} (prose, the default), {"type":"json_object"} (JSON mode, sent as response_format) or {"type":"json_schema","name":…,"schema":…,"strict":…}, which is sent as a json_schema response_format and routed only to endpoints declaring structured_outputs. Any other format.type is refused with HTTP 400 before inference.

Answer with a response.* event stream instead of one object.

Ceiling on generated tokens.

Sampling temperature.

Nucleus sampling mass.

False asks for at most one tool call per turn. Advisory: forwarded where the serving endpoint declares parallel_tool_calls, and otherwise reported unhonoured. Response objects report parallel_tool_calls: true as a placeholder.

Omit or set to null. A non-null value is refused with HTTP 400 naming previous_response_id; send the full conversation in input.

Omit or set to false/null for foreground execution. True is refused with HTTP 400 naming background.

Accepted without effect; responses are not retained server-side.

Ignored response-include hints.

Opaque request metadata; accepted and ignored.

Accepted without effect; not forwarded by the Responses translation.

An OpenAI response object, or a text/event-stream of response.* events when stream: true. Media the model generated rides in the assistant message as further content parts: an output_image part whose image_url is the picture (a data URL wherever the endpoint handed back bytes), and an output_audio part carrying data, transcript and the endpoint's own id and expires_at. Streamed, a picture is added and done in one pair of response.content_part.* events because it arrives whole, while speech accumulates through response.output_audio.delta and response.output_audio.transcript.delta and is finished by their .done events. The dialect's usage states all three counts and a client parsing it rejects an object with a member missing, so a count the endpoint never reported is sent as zero and named in usage.heimdall.unreported, which lists every member whose zero nothing measured — total_tokens among them wherever one of its terms is missing, a total short a term being unknown rather than smaller. A member absent from that list is the endpoint's own figure; the extension member is absent from a reply whose counts were all reported and names all three where no usage document arrived. response.completed carries the same object under the same rule, with honoured beside its usage where the request carried members the endpoint's declarations judged. The stream commits on its first event or after one keep-alive period, whichever comes first, so a rejection that arrives before either — an unknown model, no eligible endpoint, an engine refusing a member — is an HTTP error status with an error body carrying the public envelope's metadata. A failure after the stream has committed sends an event: error frame — {"type":"error","code":null,"message":"…","param":null}, code and param flat beside type rather than nested under an error member, matching the Responses API's own streamed error event — and then ends on response.failed, whose response.status is failed and whose response.error carries a code (invalid_prompt for a rejected request, rate_limit_exceeded, context_length_exceeded, or server_error for a failure worth retrying) and the same message, which ends with the engine's own diagnostic where one came back.

The body did not parse, or a parameter is outside the range the API accepts. A file part that names both file_id and file_data, or a request that names more than 16 stored files, is refused too. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

The key is missing, unknown, revoked, expired or past its spend limit. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

The account has no credit left to cover the request. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family. A file_id the account cannot read answers 404 as well, whether the file is another account's, was deleted or has passed its expires_at. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

The upstream did not answer within the deadline. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

The stored files the request names total more than 32 MiB. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

The marketplace failed to serve the request. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

The endpoint that was routed to failed, and no other endpoint could serve the request. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503. A request that names a stored file is also refused 503, to be retried, while the key store that stored files are sealed under cannot be reached. A request with stream: true refused before its stream opened answers in ResponsesErrorEnvelope, whose code is null; every other refusal answers in ErrorEnvelope.

Model id as author/slug, optionally suffixed :free, :nitro (fastest) or :floor (cheapest) or :exacto (measured tool-call reliability, requiring declared parameters). :online is refused with HTTP 400 because web search plugins are unsupported. provider/slug or provider/author/slug names the model that provider serves under that slug and routes to that provider alone; a catalogue id always takes precedence.

One text or a batch of them, pre-tokenized integer arrays, or a multimodal batch — items of { content: [...] } whose parts are { type: "text", text }, { type: "image_url", image_url: { url } }, or { type: "input_audio" | "input_video" | "input_file", <same name>: { data, format } } carrying the bytes base64-encoded. A multimodal input is refused on a model whose declared input modalities do not cover the parts it carries. A multimodal model can embed a content item and a plain string in two bases whose vectors are not interchangeable, so an index and the queries run against it use one shape.

Hint at how the vectors will be used, e.g. search_query, search_document.

How the vectors come back.

Truncate to this many dimensions where the model allows it.

An opaque end-user identifier.

Provider eligibility, ordering and measured performance preferences.

An OpenAI embeddings list, plus id, provider and usage.cost. usage.prompt_tokens_details carries the per-modality share of prompt_tokens (text_tokens, image_tokens, audio_tokens, video_tokens, file_tokens) and is present only when the input mixed modalities and the endpoint reported the split. Unreported token counts are omitted; explicit zeros are retained.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Model id as author/slug, optionally suffixed :free, :nitro (fastest) or :floor (cheapest) or :exacto (measured tool-call reliability, requiring declared parameters). :online is refused with HTTP 400 because web search plugins are unsupported. provider/slug or provider/author/slug names the model that provider serves under that slug and routes to that provider alone; a catalogue id always takes precedence.

What the documents are scored against.

The documents to score. Each is a string, or an object carrying text, image, or both for a vision reranker. An image document requires an endpoint declaring image input; if none qualify, routing returns 503 with input_unsupported before calling a provider or reserving credit.

How many of the highest-scoring documents come back. Absent returns all of them.

Echo each result's document back beside its score, exactly as it was submitted.

An opaque end-user identifier.

Provider eligibility, ordering and measured performance preferences.

The ranking, highest score first, plus provider and usage.cost. usage.total_tokens is omitted where the endpoint reported no count. usage.search_units counts one query against up to 100 documents as one unit, ceil(documents / 100), reported for reconciliation against the Cohere search-unit shape rather than as a charge of its own.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Model id as author/slug, optionally suffixed :free, :nitro (fastest) or :floor (cheapest) or :exacto (measured tool-call reliability, requiring declared parameters). :online is refused with HTTP 400 because web search plugins are unsupported. provider/slug or provider/author/slug names the model that provider serves under that slug and routes to that provider alone; a catalogue id always takes precedence.

The text or structured content judged by every question.

Questions keyed by caller chosen identifiers; each carries a type, instructions and any criteria.

Provider eligibility, ordering and measured performance preferences.

An opaque end-user identifier.

An opaque session identifier for routing affinity.

Trace metadata accepted by the marketplace and omitted from the upstream request.

Typed answers with probabilities and measured input tokens.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Endpoint bound to the admitted benchmark task. X-Heimdallm-Probe-Endpoint is read as the same header.

Single-use durable benchmark task identifier. X-Heimdallm-Benchmark-Task is read as the same header.

The image model to use.

What to draw.

Upper bound on images returned; an endpoint may return fewer.

Resolution tier, resolved against the endpoint's envelope into the pixels it draws.

Aspect ratio, snapped to the endpoint's own grid at the resolution tier.

A tier or explicit WIDTHxHEIGHT pixels; pixels named beside a tier or a ratio are refused with 400.

Rendering quality. Ignored by an endpoint with no such knob.

Encoding of the returned bytes. svg is vectorized markup, UTF-8 base64-encoded in b64_json.

transparent requires an output_format carrying alpha (png or webp).

Compression for webp/jpeg output. Ignored for png.

Ask for a reproducible render, where the provider supports it.

Relay the render as it is drawn, where the endpoint declares stream; the rest answer with the completed image alone.

Bounded inline image references. An https: URL parses here and is refused before dispatch: the marketplace fetches nothing on a caller's behalf.

Routing preferences, plus per-provider passthrough under options keyed by provider slug.

An opaque end-user identifier, folded into a hashed id before it reaches a provider.

The paid artifact is retained for scoring.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

Benchmark dispatch is unavailable to this caller.

The upstream did not answer within the deadline.

The task is not eligible for dispatch.

The request body is larger than this route accepts. The inline assets a request carries are bounded by their own decoded-byte budget, which is read only once a body has arrived; a body over the route’s size limit is refused here instead.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Endpoint bound to the admitted benchmark task. X-Heimdallm-Probe-Endpoint is read as the same header.

Single-use durable benchmark task identifier. X-Heimdallm-Benchmark-Task is read as the same header.

The paid artifact is retained for scoring.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

Benchmark dispatch is unavailable to this caller.

The upstream did not answer within the deadline.

The task is not eligible for dispatch.

The request body is larger than this route accepts. The inline assets a request carries are bounded by their own decoded-byte budget, which is read only once a body has arrived; a body over the route’s size limit is refused here instead.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Endpoint bound to the admitted benchmark task. X-Heimdallm-Probe-Endpoint is read as the same header.

Single-use durable benchmark task identifier. X-Heimdallm-Benchmark-Task is read as the same header.

The paid artifact is retained for scoring.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

Benchmark dispatch is unavailable to this caller.

The upstream did not answer within the deadline.

The task is not eligible for dispatch.

The request body is larger than this route accepts. The inline assets a request carries are bounded by their own decoded-byte budget, which is read only once a body has arrived; a body over the route’s size limit is refused here instead.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Endpoint bound to the admitted benchmark task. X-Heimdallm-Probe-Endpoint is read as the same header.

Single-use durable benchmark task identifier. X-Heimdallm-Benchmark-Task is read as the same header.

Model id as author/slug, optionally suffixed :free, :nitro (fastest) or :floor (cheapest) or :exacto (measured tool-call reliability, requiring declared parameters). :online is refused with HTTP 400 because web search plugins are unsupported. provider/slug or provider/author/slug names the model that provider serves under that slug and routes to that provider alone; a catalogue id always takes precedence.

The text to synthesize. Its length in characters is the unit billed.

A voice an endpoint of the model declares, which GET /speech/models publishes. Absent asks for whichever voice the endpoint defaults to. The reference contract's own names (alloy, ash, ballad, coral, echo, fable, nova, onyx, sage, shimmer, verse) are accepted wherever an endpoint declares them, and where none does the request resolves to the one voice the model's endpoints offer between them — named back in X-Heimdall-Voice — or answers 400 listing the voices it is served with.

The codec the audio comes back in. Absent is mp3.

Playback rate, honoured only where the endpoint declares a range covering it.

Sampling seed, honoured only where the endpoint declares it.

Generation token limit, honoured only where the endpoint declares it.

Free text steering delivery, where the endpoint honours it.

One voice sample to clone, at most 5 MiB of decoded bytes. Requires declared speech input_references and audio input. The sample takes precedence over voice and is not saved as a voice profile.

An opaque end-user identifier.

Provider eligibility, ordering and measured performance preferences.

The paid artifact is retained for scoring.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

Benchmark dispatch is unavailable to this caller.

The upstream did not answer within the deadline.

The task is not eligible for dispatch.

The request body is larger than this route accepts. The inline assets a request carries are bounded by their own decoded-byte budget, which is read only once a body has arrived; a body over the route’s size limit is refused here instead.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

The image model to use.

What to draw.

Upper bound on images returned; an endpoint may return fewer.

Resolution tier, resolved against the endpoint's envelope into the pixels it draws.

Aspect ratio, snapped to the endpoint's own grid at the resolution tier.

A tier or explicit WIDTHxHEIGHT pixels; pixels named beside a tier or a ratio are refused with 400.

Rendering quality. Ignored by an endpoint with no such knob.

Encoding of the returned bytes. svg is vectorized markup, UTF-8 base64-encoded in b64_json.

transparent requires an output_format carrying alpha (png or webp).

Compression for webp/jpeg output. Ignored for png.

Ask for a reproducible render, where the provider supports it.

Relay the render as it is drawn, where the endpoint declares stream; the rest answer with the completed image alone.

Bounded inline image references. An https: URL parses here and is refused before dispatch: the marketplace fetches nothing on a caller's behalf.

Routing preferences, plus per-provider passthrough under options keyed by provider slug.

An opaque end-user identifier, folded into a hashed id before it reaches a provider.

The rendered images and what they cost, or the stream that drew them.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

The request body is larger than this route accepts. The inline assets a request carries are bounded by their own decoded-byte budget, which is read only once a body has arrived; a body over the route’s size limit is refused here instead.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Frames per second, restricted to the serving endpoint’s frame grid.

Content the clip should avoid.

Denoising steps, within the endpoint’s declared range.

Prompt adherence, within the endpoint’s declared range.

How closely to follow the supplied reference.

A sampler the endpoint declares.

A scheduler the endpoint declares.

The model to render with, as the catalogue spells it.

What the clip should show. Optional: a model that animates a supplied frame takes the picture in place of the words, and one that needs a prompt refuses a request without one.

Requested clip length in seconds. It is the meter: the bill is the delivered duration at the endpoint’s declared per-second rate.

The resolution tier to render at, narrowed per model by supported_resolutions. Where an endpoint prices its tiers separately, this is what selects the rate.

The shape to render at, narrowed per model by supported_aspect_ratios.

Exact pixel dimensions as WIDTHxHEIGHT, interchangeable with a resolution tier and an aspect ratio.

Fixing it makes a generation reproducible on the same endpoint.

Whether to score a soundtrack into the clip. Omission leaves the choice to the endpoint’s own default. Sound is billed by the same per-second rate as the rest of the clip, so asking for it adds no second charge.

Images to pin to the first and last frame. The role is on the image rather than on its position in the array, because an array of one carries no position that says which end it is.

Assets that condition the generation without being frames of it, each carrying what it is for as role: reference when absent, or a control pass the render keeps the structure of. An image is never a frame here — a pinned still is a frame_images item, and reaches the endpoint under first_frame or last_frame. A role is admitted per kind, as GET /videos/models lists under supported_reference_roles: a picture or a clip takes any control pass, and a sample takes reference alone. The job is routed only to an endpoint declaring every role the array carries for that kind, and a combination of control passes no endpoint loads together fails it naming the sets that are.

Where to post the job once it reaches a terminal status, so a caller need not poll. HTTPS only: the notification names a job an API key can read, and a cleartext hop would publish it.

Routing preferences, plus the per-slug passthrough carrying an engine’s own knobs. Only the keys GET /videos/models publishes under allowed_passthrough_parameters are forwarded.

The job was accepted. status is where it starts and polling_url is where to read it next; neither is a claim that anything has been rendered or charged yet.

The body did not parse, a parameter is outside the range the API accepts — a resolution, aspect ratio or frame role outside the contract’s own enum, a size that is not WIDTHxHEIGHT, an asset that is not an inline base64 data URI, or a callback_url that is not HTTPS — or the assets the job names cannot be served as asked. frame_images and input_references are folded into one list before dispatch, so the two together carry at most 3 items and at most 20 MiB of decoded bytes; a job over either bound is refused naming it. A combination of control roles no endpoint of the model loads in one render is refused naming the sets it does serve, which reference_inputs.video.control_sets on GET /models/{author}/{slug}/endpoints lists.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

The request body is larger than this route accepts. The inline assets a request carries are bounded by their own decoded-byte budget, which is read only once a body has arrived; a body over the route’s size limit is refused here instead.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

The job id POST /videos returned, of the form vjob_ followed by 32 hex characters.

The job as it stands.

The key is missing, unknown, revoked, expired or past its spend limit.

No such job on the key’s own account. Identical whether the id is unknown, malformed, or names a job another account submitted.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The job id POST /videos returned, of the form vjob_ followed by 32 hex characters.

Which clip of a job that returned several. The first one by default.

One byte range, as bytes=first-last, bytes=first- or bytes=-count; the unit is read in any case. A response that starts at the first byte and is written through the last is a delivery, so bytes=0- delivers like a whole fetch and a range starting later never does. Any other form, several ranges included, is ignored and the whole asset is served.

The clip, as the container the provider produced it in. The body is the video itself, with no JSON envelope.

The requested range of the asset. It is not a delivery and removes nothing.

index is not a whole number at or above zero.

The key is missing, unknown, revoked, expired or past its spend limit.

No such job on the key’s own account, or the job produced no clip at that index.

The job has not completed, so there is no content to serve yet.

The clip is gone: its grace after delivery ended, or it went undelivered until the job expired.

The range starts past the final byte. Content-Range states the asset’s length as bytes */size.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

Optional marketplace Blender repair with a selected vision model. Generation, model calls, and machine time are billed separately.

The model to generate with, as the catalogue spells it.

What to build. Optional with a supplied view; an image-only engine refuses text alongside a view.

The most faces the mesh may carry, and a ceiling rather than a target: a decimator lands at or below it. Bounded here only against a typo; what an endpoint generates is its mesh_envelope, which GET /meshes/models publishes, and a count outside it is an unrouted request.

Shape cascade resolution in voxels along each axis, within the selected endpoint’s declared choices.

Square texture atlas edge in pixels, within the selected endpoint’s declared choices.

Horizontal camera field of view in degrees, where the selected endpoint declares it.

Whether to bake a texture into the mesh. Omission leaves the choice to the endpoint’s own default. A texture is billed by the declared per-mesh rate like the rest of the asset unless the endpoint declares a tier for it, so asking for it adds no second charge.

Views of one object to build from, at most four, because more views are a different rig and a different model. Inline only: the marketplace fetches nothing on this rail, so a URL it accepted would be an ask it could not carry out.

Fixing it makes a generation reproducible on the same endpoint.

What the text-to-image stage should avoid.

Sampling steps, within the provider’s declared limits.

Conditioning strength, within the selected endpoint’s declared limits.

A sampling algorithm supported by the selected endpoint.

A sampling schedule supported by the selected endpoint.

Where to post the job once it reaches a terminal status, so a caller need not poll. HTTPS only: the notification names a job an API key can read, and a cleartext hop would publish it.

Routing preferences, plus the per-slug passthrough carrying an engine’s own knobs. Only the keys GET /meshes/models publishes under allowed_passthrough_parameters are forwarded.

The job was accepted. status is where it starts and polling_url is where to read it next; neither is a claim that anything has been generated or charged yet.

The body did not parse, or a parameter is outside the range the API accepts — a request naming neither a prompt nor a view, a target_faces outside the contract’s own bounds, more than four views, or a callback_url that is not HTTPS.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

The job id POST /meshes returned, of the form mjob_ followed by 32 hex characters.

The job as it stands.

The key is missing, unknown, revoked, expired or past its spend limit.

No such job on the key’s own account. Identical whether the id is unknown, malformed, or names a job another account submitted.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The job id POST /meshes returned, of the form mjob_ followed by 32 hex characters.

Which asset of a job that returned several. The first one by default.

One byte range, as bytes=first-last, bytes=first- or bytes=-count; the unit is read in any case. A response that starts at the first byte and is written through the last is a delivery, so bytes=0- delivers like a whole fetch and a range starting later never does. Any other form, several ranges included, is ignored and the whole asset is served.

The mesh, as a binary glTF container. The body is the asset itself, with no JSON envelope.

The requested range of the asset. It is not a delivery and removes nothing.

index is not a whole number at or above zero.

The key is missing, unknown, revoked, expired or past its spend limit.

No such job on the key’s own account, or the job produced no mesh at that index.

The job has not completed, so there is no content to serve yet.

The mesh is gone: its grace after delivery ended, or it went undelivered until the job expired.

The range starts past the final byte. Content-Range states the asset’s length as bytes */size.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The job id POST /meshes returned, of the form mjob_ followed by 32 hex characters.

Cancellation request recorded.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Deployment availability, exact machine price, and live vision models with tool support.

Too many requests

The encoded audio, with the codec's own content type: audio/mpeg for mp3, audio/wav for wav, audio/pcm for pcm, audio/ogg; codecs=opus for opus, audio/flac for flac, audio/aac for aac, audio/basic for mulaw and audio/x-alaw-basic for alaw. No Content-Length: an endpoint that encodes as it decodes has no length to declare.

The body did not parse, a parameter is outside the range the API accepts, or voice names a voice no endpoint of the model declares and the model's endpoints offer no single substitute — that error lists the voices the model is served with under metadata.declared_voices.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Model id as author/slug, optionally suffixed :free, :nitro (fastest) or :floor (cheapest) or :exacto (measured tool-call reliability, requiring declared parameters). :online is refused with HTTP 400 because web search plugins are unsupported. provider/slug or provider/author/slug names the model that provider serves under that slug and routes to that provider alone; a catalogue id always takes precedence.

Base64-encoded audio, on the JSON body variant only.

ISO-639-1 code of the spoken language, where known.

Context to bias spelling and punctuation.

What the transcript comes back as.

Transcript generation token limit, where the endpoint declares it.

Sampling temperature.

Repeat the field as timestamp_granularities[] for each level wanted.

Refused on this route: transcriptions do not stream.

Provider eligibility, ordering and measured performance preferences.

response_format: verbose_json adds task, duration, language, segments and words (segments and words carry speaker where the endpoint returns diarization). usage.seconds is whole seconds, rounded up from what the endpoint measured, which is the unit a duration price is charged in; duration is that measurement itself and stays fractional. Unreported usage figures are omitted; an explicitly measured zero is retained.

The body did not parse, or a parameter is outside the range the API accepts.

The key is missing, unknown, revoked, expired or past its spend limit.

The account has no credit left to cover the request.

No such model. When the id names a family whose base model is not served, metadata.family_variants lists the served ids in that family.

The upstream did not answer within the deadline.

The audio file is larger than the 25 MiB the multipart body accepts.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The endpoint that was routed to failed, and no other endpoint could serve the request.

No live endpoint can serve the request: none is declared for the model, or every one of them was excluded by something the request asked for. A 503 that carries Retry-After and X-Should-Retry: true is the marketplace restarting, and repeating the request is correct. The same restart reaches a stream that has already begun as an in-band error carrying 503.

Records to skip.

Maximum records to return.

The served embeddings models.

offset or limit is outside the range this route accepts.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

Records to skip.

Maximum records to return.

The served image models.

offset or limit is outside the range this route accepts.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the model id before the slash.

The half of the model id after the slash.

The model id and its live image endpoints.

No such image model.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

Records to skip.

Maximum records to return.

The served video models.

offset or limit is outside the range this route accepts.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

Records to skip.

Maximum records to return.

The served mesh models.

offset or limit is outside the range this route accepts.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

Records to skip.

Maximum records to return.

The served speech models.

offset or limit is outside the range this route accepts.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

Content validators from earlier responses; a current match returns 304 without a body.

The complete public catalogue. Array entries have the corresponding list route shapes.

The current representation matches If-None-Match; no response body.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The catalogue is unavailable or exceeds the response size limit.

Restrict the list to one family id.

Case-insensitive substring match against the model id or its display name.

Unsupported: this catalogue computes no curated usage category. Reported in unsupported_filters.

Comma-separated parameter names. A model matches when every named parameter is declared by at least one of its live endpoints, not necessarily the same one.

Comma-separated modalities the model must accept, all of them (text, image, audio, video, file).

Comma-separated modalities the model must produce, all of them (text, image, video, audio, speech, transcription, embeddings, rerank, decisions, mesh).

Minimum context length in tokens, against the model's own declared window.

Minimum prompt price. At least one live endpoint must price at or above this. A decimal USD-per-million-token string such as "1.50", compared exactly against the endpoint's declared per-token price. Never a float. An endpoint publishing no rate per token satisfies neither bound.

Maximum prompt price. At least one live endpoint must price at or below this. A decimal USD-per-million-token string such as "1.50", compared exactly against the endpoint's declared per-token price. Never a float. An endpoint publishing no rate per token satisfies neither bound.

Minimum completion price. At least one live endpoint must price at or above this. A decimal USD-per-million-token string such as "1.50", compared exactly against the endpoint's declared per-token price. Never a float. An endpoint publishing no rate per token satisfies neither bound.

Maximum completion price. At least one live endpoint must price at or below this. A decimal USD-per-million-token string such as "1.50", compared exactly against the endpoint's declared per-token price. Never a float. An endpoint publishing no rate per token satisfies neither bound.

Comma-separated provider slugs. At least one live endpoint must belong to one of them.

Comma-separated authors, matched against the model's own declared author.

true restricts the list to models with at least one endpoint the ZDR fold keeps (see GET /endpoints/zdr). false is accepted and filters nothing.

Comma-separated country codes or datacenter regions. At least one live endpoint must declare a matching datacenter.

Unsupported: this catalogue declares no architecture classification beyond modality. Reported in unsupported_filters.

Unsupported: this catalogue declares no distillation licence flag. Reported in unsupported_filters.

Order of the returned list. newest (default) and oldest order by release date; the rest need a computed figure this catalogue has.

Minimum age in days since the model's declared release date.

Maximum age in days since the model's declared release date.

Minimum measured tool_call_success_1d, 0-100. At least one live endpoint must clear its sample gate and meet the bound; a null figure matches neither bound.

Maximum measured tool_call_success_1d, 0-100. At least one live endpoint must clear its sample gate and meet the bound; a null figure matches neither bound.

Unsupported: this catalogue computes no intelligence index. Reported in unsupported_filters.

Unsupported: this catalogue computes no intelligence index. Reported in unsupported_filters.

Unsupported: this catalogue computes no coding index. Reported in unsupported_filters.

Unsupported: this catalogue computes no coding index. Reported in unsupported_filters.

Unsupported: this catalogue computes no agentic index. Reported in unsupported_filters.

Unsupported: this catalogue computes no agentic index. Reported in unsupported_filters.

Page size. Without it the whole list is returned.

Rows to skip. Valid on its own, without limit.

The served catalogue, filtered and paginated.

A parameter is missing, repeated, or outside the range or enum this route accepts.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The number of served models.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The account's own model list.

The key is missing, unknown, revoked, expired or past its spend limit.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the model id before the slash.

The half of the model id after the slash.

The served model.

No such served model.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the model id before the slash.

The half of the model id after the slash.

The served model.

No such served model.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The whole model id, percent-encoded as one path segment (author%2Fslug) rather than split across two.

The served model, as a bare object.

No such served model.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the model id before the slash.

The half of the model id after the slash.

The model shell, its family and its live endpoints.

No such model. When the id names a family whose base model is not served, metadata.family and metadata.family_variants name the served ids to redirect to.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the model id before the slash.

The half of the model id after the slash.

Validators from earlier responses; a current match returns 304 without a body.

The publishable scores.

The body is the one If-None-Match names; no response body.

No such model.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the model id before the slash.

The half of the model id after the slash.

Days of history, ending now.

Validators from earlier responses; a current match returns 304 without a body.

One series per live endpoint.

The body is the one If-None-Match names; no response body.

days is outside 1 to 7.

No such model.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the model id before the slash.

The half of the model id after the slash.

Days of history, ending today.

Validators from earlier responses; a current match returns 304 without a body.

The daily counts.

The body is the one If-None-Match names; no response body.

days is outside 1 to 30.

No such model.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

Every family with a served model in it.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the family id before the slash.

The half of the family id after the slash.

The family, its served models newest first, and their live endpoints.

No such family.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the family id before the slash.

The half of the family id after the slash.

Days of history, ending now.

Validators from earlier responses; a current match returns 304 without a body.

One series per live endpoint in the family.

The body is the one If-None-Match names; no response body.

days is outside 1 to 7.

No such family.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The half of the family id before the slash.

The half of the family id after the slash.

Days of history, ending today.

Validators from earlier responses; a current match returns 304 without a body.

The daily counts across the family.

The body is the one If-None-Match names; no response body.

days is outside 1 to 30.

No such family.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The approved providers.

Too many requests for the key, or for the address when no key was sent. Retry-After says when to try again.

The marketplace failed to serve the request.

The provider slug.